Object detectors such as OWLv2 return many overlapping boxes for the same object. Intersection over Union measures how much two boxes overlap, and Non-Maximum Suppression uses it to keep one box per object.

Intersection Over Union

For boxes given as corners :

  • intersection width , and the same for height
  • IoU is 0 for boxes that do not touch and 1 for identical boxes
  • it also scores detections against ground truth, where IoU of at least 0.5 often counts as correct

Non-Maximum Suppression

Steps:

  • drop boxes below a score threshold (for example 0.3)
  • sort the rest by score, highest first
  • keep the top box, then remove every remaining box whose IoU with it is at least a threshold (for example 0.5)
  • repeat with the next highest remaining box

Choices:

  • class-aware NMS only lets boxes of the same label suppress each other
  • the IoU threshold trades duplicates (too high) against losing objects that really overlap (too low)
  • soft-NMS lowers the scores of overlapping boxes instead of deleting them, which helps in crowded scenes

Top-1 Per Query

  • NMS cannot reduce detections to one per label, because boxes on two different objects do not overlap
  • when a scene is known to hold at most one of each object, keeping only the best box per label does that job
  • turn it off as soon as a scene can hold two objects with the same label

Keep Filtering Separate From The Model

  • return every raw box from the detector and filter afterwards
  • the filter code can then be unit tested without the model, and the number of boxes after each step can be logged